A cheap spiking neural model for evolved controllers

Abstract

When implementing spiking neural networks into agents acting in virtual environments, there often exist an underlying problem of incongruity between the level of detail in the (biological) neural model and the agent’s environment, resulting in excessive network iterations required for functional behavior. This thesis aims to explore a new spiking neural model specifically made to address this issue, and will be assessed by its similarity to biological neurons, its cost in computation, and finally its behavioral capabilities when acting as a controller for a virtual creature. The results show that our new spiking model is significantly cheaper to implement, in addition to outperform other neural models in certain behavioral tests.